US20090157400A1

US20090157400A1 - Speech recognition system and method with cepstral noise subtraction

Info

Publication number: US20090157400A1
Application number: US12/243,303
Authority: US
Inventors: Shih-Ming Huang
Original assignee: Industrial Technology Research Institute ITRI
Current assignee: Industrial Technology Research Institute ITRI
Priority date: 2007-12-14
Filing date: 2008-10-01
Publication date: 2009-06-18
Also published as: JP5339426B2; JP2009145895A; TW200926141A; TWI356399B; US8150690B2

Abstract

The invention relates to a speech recognition system and method with cepstral noise subtraction. The speech recognition system and method utilize a first scalar coefficient, a second scalar coefficient, and a determining condition to limit the process for the cepstral feature vector, so as to avoid excessive enhancement or subtraction in the cepstral feature vector, so that the operation of the cepstral feature vector is performed properly to improve the anti-noise ability in speech recognition. Furthermore, the speech recognition system and method can be applied in any environment, and have a low complexity and can be easily integrated into other systems, so as to provide the user with a more reliable and stable speech recognition result.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention
The present invention relates to a speech recognition system and method, more particularly to a speech recognition system and method with cepstral noise subtraction.
2. Description of the Related Art
Speech is the most direct method of communication for human beings, and computers used in daily life also have a speech recognition function. For example, the Windows XP operating system of Microsoft provides this function, and so does the latest Windows Vista operating system. Also, the latest operating system Mac OS X of another company, Apple, provides a speech recognition function.
No matter whether a microphone is used to carry out the speech recognition function on a computer using Microsoft Windows XP/Vista or Apple Mac OS X or a phone call is made through the service provided by Google and Microsoft, the speech will be processed by an electronic device such as a microphone or a telephone, which may interfere with the voice signal. Also, other background noises, e.g., sounds made by air conditioners or people walking, may also greatly reduce the speech recognition rate. Therefore, a good anti-noise speech recognition technique is in high demand.
The conventional cepstral mean subtraction (CMS) used for speech recognition (see paper [1] in the prior art Furui, “Cepstral analysis technique for automatic speaker verification,” IEEE Transaction on Acoustics, Speech and Signal Processing, 29, pp. 254-272, 1981.) has become a widely used feature processing method for enhancing the anti-noise ability in speech recognition.
U.S. Pat. No. 6,804,643 has also disclosed a cepstral feature processing method as shown in FIG. 1. In Step S11, first cepstral mean vectors of all the voice frames before the current voice frame are first calculated. In Step S12, a sampling value is then received, i.e., the cepstral feature vector of the current voice frame is used. In Step S13, the cepstral feature vector of the current voice frame has an estimated mean vector added. The estimated mean vector is an adjustment factor multiplied by a cepstral mean vector of the preceding voice frame. In Step S14, a new estimated cepstral feature vector is calculated.
Therefore, it is necessary to provide a speech recognition system with cepstral noise subtraction to improve the function of anti-noise speech recognition.

SUMMARY OF THE INVENTION

The present invention provides a speech recognition system with cepstral noise subtraction which includes a filterbank energy extractor, a cepstral noise subtraction device, a cepstral converter, a model trainer, and a speech recognizer. The filterbank energy extractor obtains a plurality of first feature vectors according to a voice signal. The cepstral noise subtraction device obtains a first feature vector of a preset voice frame and first feature vectors of a plurality of voice frames before the preset voice frame, so as to calculate a feature mean vector, and calculate a second feature vector of a preset voice frame according to the first feature vector, the feature mean vector, a first scalar coefficient, and a second scalar coefficient of the preset voice frame. The cepstral converter converts the second feature vector of the preset voice frame into a cepstral feature vector. The model trainer calculates a model parameter according to the cepstral feature vector. The speech recognizer calculates a recognized voice signal according to the cepstral feature vector and the model parameter.
The present invention provides a speech recognition method with cepstral noise subtraction which includes the following steps. A plurality of first feature vectors is obtained according to a voice signal. A first feature vector of a preset voice frame and first feature vectors of a plurality of voice frames before the preset voice frame are obtained to calculate a feature mean vector. A second feature vector of a preset voice frame is calculated according to a first feature vector, the feature mean vector, a first scalar coefficient, and a second scalar coefficient of the preset voice frame. The second feature vector of the preset voice frame is converted into a cepstral feature vector. A model parameter is calculated according to the cepstral feature vector. A recognized voice signal is calculated according to the cepstral feature vector and the model parameter.
According to the speech recognition system and method of the present invention, the process for the cepstral feature vector is limited, so as to avoid excessive enhancement and subtraction in the cepstral feature vector, so that its operation is completed properly, and the anti-noise ability for speech recognition is improved. Furthermore, the speech recognition system and method can be applied in any environment, and have a low complexity and can be easily integrated into other systems, so as to provide the user with a more reliable and stable speech recognition result.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic flow chart of a conventional cepstral feature processing method;

FIG. 2 is a schematic block diagram of a speech recognition system with cepstral noise subtraction according to the present invention;

FIG. 3 is a schematic flow chart of the cepstral noise subtraction method according to the present invention;

FIG. 4 is a schematic block diagram of the cepstral noise subtraction device according to the present invention;

FIG. 5 is a schematic flow chart of the calculation of a feature mean vector according to the present invention; and

FIG. 6 is a schematic block diagram of a feature mean vector calculator device according to the present invention.

DETAILED DESCRIPTION OF THE INVENTION

FIG. 2 is a schematic block diagram of a speech recognition system with cepstral noise subtraction according to the present invention. According to the present invention, the speech recognition system 20 with cepstral noise subtraction includes a filterbank energy extractor 21, a cepstral noise subtraction device 22, a cepstral converter 23, a model trainer 25, and a speech recognizer 27. The filterbank energy extractor 21 obtains a plurality of first feature vectors according to a voice signal. In this embodiment, the filterbank energy extractor 21 is a log Mel filterbank energy extractor. By the use of the log Mel filterbank energy extractor, the first feature vectors are log Mel filterbank energy feature vectors.
The cepstral noise subtraction device 22 obtains a first feature vector of a preset voice frame and first feature vectors of a plurality of voice frames before the preset voice frame, so as to calculate a feature mean vector, and calculate a second feature vector of a preset voice frame according to the first feature vector, the feature mean vector, a first scalar coefficient, and a second scalar coefficient of the preset voice frame.
FIG. 4 is a schematic block diagram of the cepstral noise subtraction device according to the present invention. The cepstral noise subtraction device 22 of the present invention includes a feature mean vector calculator device 41, a first multiplier 42, a first adder 43, a second multiplier 44, a comparator 45, and a multiplexer 46. The feature mean vector calculator device 41 obtains the first feature vector of the preset voice frame and the first feature vectors of the plurality of voice frames before the preset voice frame, so as to calculate the feature mean vector.
In this embodiment, the number of the plurality of voice frames before the preset voice frame is between 2 and a total number of voice frames of a sentence. If the total number of the voice frames of a sentence is N, the feature mean vector calculator device 41 obtains the first feature vector of the N voice frames before the preset voice frame, and calculates the feature mean vector, which is expressed by Formula (1) below:
$\begin{matrix} \overline{X} = \frac{1}{N} (X_{t - (N - 1)} + \dots + X_{t - 2} + X_{t - 1} + X_{t}) & (1) \end{matrix}$
where X_tis the first feature vector of the preset voice frame, X_t-1to X_t-(N-1)are the first feature vectors of the plurality of voice frames before the preset voice frame, N is the number of the voice frames, and X is the feature mean vector.
FIG. 6 is a schematic block diagram of the feature mean vector calculator device according to the present invention. The feature mean vector calculator device 41 of the present invention includes a plurality of delayers 411, 412, 415, a second adder 416, and a third multiplier 417. Each delayer delays a unit of time, so as to obtain the first feature vectors of the plurality of voice frames before the preset voice frame. The second adder 416 sums the first feature vectors, so as to calculate a sum of the first feature vectors (X_t-(N-1)+ . . . +X_t-2+X_t-1+X_t). The third multiplier 417 multiplies the sum of the first feature vectors (X_t-(N-1)+ . . . +X_t-2+X_t-1+X_t) with a reciprocal (1/N) of the number of the voice frames, so as to calculate the feature mean vector X.
FIG. 5 is a schematic flow chart of the calculation of the feature mean vector according to the present invention. First, in Step S52, a parameter Temp is set as a zero vector. In Step S53, a parameter p is set as zero, where the p indicates the p^thvoice frame. In Step S54, the first feature vectors of the preset voice frames are summed to calculate a sum of the first feature vectors. In Steps S55 and S56, whether the p^thvoice frame has reached N-1 or not is determined. If negative, p is incremented. The step of adding p is the above step of using a delayer to delay a unit of time, so as to obtain the first feature vectors of the plurality of voice frames before the preset voice frame. In Step S57, if the p has reached the number of N-1, the sum of the first feature vectors (Temp) is multiplied with the reciprocal (1/N) of the number of the voice frames. In Step S58, the feature mean vector X is calculated.
In the above embodiment, the feature mean vector is calculated through an arithmetic mean. However, in the feature mean vector calculator device and method of the present invention, the mean calculation methods including geometric mean, median, mode, or norm may also be used to calculate the feature mean vector.
In FIG. 4, after the feature mean vector calculator device 41 calculates the feature mean vector, the first multiplier 42 multiplies the feature mean vector ( X) by the negative value (−α) of the first scalar coefficient to calculate a first multiplication result (−α· X). The first adder 43 adds the first feature vector (X_t) of the preset voice frame with the first multiplication result (−α· X) to calculate an addition result (X_t−α· X). The second multiplier 44 multiplies the first feature vector (X_t) of the preset voice frame by the second scalar coefficient (β) to calculate a second multiplication result (β·X_t). The comparator 45 compares whether the addition result (X_t−α· X) is greater than the second multiplication result (β·X_t), and outputs a control signal to the multiplexer 46. The multiplexer 46 switches the second feature vector ({circumflex over (X)}_t) of the preset voice frame into the addition result (X_t−α· X) or the second multiplication result (β·X_t) according to the control signal.
Therefore, in the system and method of the present invention, after the cepstral noise subtraction device 22 calculates the feature mean vector, the feature vector and the feature mean vector of the preset voice frame are operated under certain conditions, which is expressed by Formula (2):
$\begin{matrix} {\hat{X}}_{t} = {\begin{matrix} X_{t} - α \cdot \overline{X} & if X_{t} > \frac{α}{1 - β} \cdot \overline{X} \\ β \cdot X_{t} & otherwise \end{matrix} & (2) \end{matrix}$
where, when the addition result (X_t−α· X) is greater than the second multiplication result (β·X_t), the second feature vector ({circumflex over (X)}_t) of the preset voice frame is the addition result (X_t−α· X), and when the addition result (X_t−α· X) is smaller than the second multiplication result (β·X_t), the second feature vector ({circumflex over (X)}_t) of the preset voice frame is the second multiplication result (β·X_t). Moreover, the first scalar coefficient (α) is between 0.01 and 0.99, and the second scalar coefficient (β) is between 0.01 and 0.99.
FIG. 3 is a schematic flow chart of the cepstral noise subtraction method according to the present invention. First, in Step S31, a parameter n is set as 1, where n indicates the n^thvoice frame, and the input speech is assumed to have L voice frames in this embodiment. In Step S32, the feature mean vector is calculated, which may refer to the description of FIGS. 5 and 6, and will not be repeated herein. Thus, the first feature vector of the preset voice frame (n) and the first feature vectors of the plurality of voice frames before the preset voice frame are obtained to calculate the feature mean vector. Then the feature mean vector ( X) is multiplied by the negative value (−α) of the first scalar coefficient to calculate a first multiplication result (−α· X). Then the first feature vector (X_t) of the preset voice frame is added to the first multiplication result (−α· X) to calculate the addition result (X_t−α· X). Then, the first feature vector (X_t) of the preset voice frame is multiplied by the second scalar coefficient (β) to calculate a second multiplication result (β·X_t).
In Step S33, whether a condition A is true or not is determined. The condition A is the condition in the above Formula (2), i.e., whether the addition result (X_t−α· X) is greater than the second multiplication result (β·X_t). In Step S34, when the addition result (X_t−α· X) is greater than the second multiplication result (β·X_t), a first operation is performed to make the second feature vector ({circumflex over (X)}_t) of the preset voice frame into the addition result (X_t−α· X). In Step S35, when the addition result (X_t−α· X) is smaller than the second multiplication result (β·X_t), a second operation is performed to make the second feature vector ({circumflex over (X)}_t) of the preset voice frame into the second multiplication result (β·X_t). In Step S36, the second feature vector ({circumflex over (X)}_t) of the preset voice frame is calculated through the above operations.
In Steps S37 and S38, if the input speech in this embodiment is assumed to have L voice frames, the calculation should be performed L times to determine whether the preset voice frame (n) has reached L; if negative, n is incremented. In Step S39, the second feature vectors ({circumflex over (X)}_t) of all voice frames are calculated.
In FIG. 2, the cepstral converter 23 converts the second feature vector of the preset voice frame into a cepstral feature vector. In this embodiment, the cepstral converter 23 is a discrete cosine transformer, and the cepstral feature vector is a Mel cepstral feature vector. The model trainer 25 calculates a model parameter according to the cepstral feature vector. The speech recognizer 27 calculates the recognized voice signal according to the cepstral feature vector and the model parameter.
The speech recognition system 20 with cepstral noise subtraction of the present invention further includes a differential operator 24 for calculating a first-order difference, or a first-order difference and a second-order difference, or a first-order difference to a high-order difference of the cepstral feature vector. In FIG. 2, the speech passes through the filterbank energy extractor 21, the cepstral noise subtraction device 22, the cepstral converter 23, the differential operator 24, and the speech recognizer 27, and thus, the recognized voice signal is calculated. The right side of the dashed line is referred to as a recognition phase. At the left side of the dashed line, the process through the model trainer 25 and a speech model parameter database 26 is referred to as a training phase. The differential operator 24 may be disposed in the recognition phase or the training phase to perform a difference operation.
The system and method of the present invention conduct experiments under the international standard Aurora-2 speech database environment to evaluate the anti-noise ability. The speech database Aurora-2 used in the experiment is issued by the European Telecommunications Standards Institute (ESTI), and is a consecutive English number speech containing noise. The noise includes eight different kinds of additive noises and two channel effects with different characteristics. The additive noise in the speech database includes airport, babble, car, exhibition, restaurant, subway, street and train station, which are added to clean speech according to different signal-to-noise ratios (SNR). The SNR includes 20 dB, 15 dB, 10 dB, 5 dB, 0 dB, and −5 dB. The channel effect includes two standards,—G.712 and MIRS, established by the International Telecommunication Union (ITU). According to different types of channel noise and additive noise added to the test speech, the Aurora-2 is divided into three test groups, Set A, Set B, and Set C. Set A represents stationary noises, and Set B represents nonstationary noises. Besides the stationary and nonstationary noise, Set C further uses the channel effects G.712 and MIRS that are different from the training speech. The average recognition rate in all kinds of noises is obtained by calculating the average value of 20 dB to 0 dB.
The speech recognition experiment is used together with an HTK (Hidden Markov Model Toolkit) development tool. The HTK is a hidden Markov model (HMM) developed by the electrical mechanism department in Cambridge University. Thus, a speech recognition system with an HMM architecture may be developed conveniently and quickly.
The settings of the acoustic models are described as follows. Each number model (1-9, zero, and oh) is modeled by a continuous density hidden Markov model (CDHMM) in a left-to-right form, and includes 16 states. Each state is modeled by three Gaussian mixture distributions. Moreover, the silence model includes two models, namely a silence model including three states indicating the silence at the beginning and the end of a sentence, and a pause model including 6 states indicating a short intermittence between words in the sentence. All the above training of the acoustic models and all the experiments are accomplished in the Aurora-2 speech database environment working together with the HTK tool suit.
As for the feature extractor, the evaluation experiment on the system and method of the present invention employs the Mel-frequency cepstral coefficients (MFCCs) as the speech feature vectors. The system and method of the present invention perform operations on log Mel filterbank energy excluding the log energy. The log Mel filterbank energy and the Mel-frequency cepstral coefficient are in a linear conversion relationship, and thus, the two are equivalent to each other. The voice frame length is sampled at 25 ms, and the voice frame shift is 10 ms. The information of each voice frame is indicated by 39-dimension, including 12-dimension Mel-frequency cepstral coefficient and 1-dimension log energy. Meanwhile, the first-order difference coefficient (delta coefficient) and the second-order difference coefficient (acceleration coefficient) corresponding to the 13-dimension feature are selected.
The recognition result is shown in FIG. 1. Compared with the cepstral mean substraction (CMS) and the prior American patent (U.S. Pat. No. 6,804,643 B1), the system and method of the present invention have obviously improved word accuracy, and the maximum word accuracy is shown in bold. As for the overall performance of set A, set B, and set C, the system and method of the present invention may effectively improve the anti-noise speech recognition rate, and are also proved to be stable and effective.
The speech recognition system and method limit the process for the cepstral feature vector, so as to avoid excessive enhancement and subtraction in the cepstral feature vector, so that its operation is performed properly to improve anti-noise ability in speech recognition. Furthermore, the speech recognition system and method can be applied in any environment, and have a low complexity and can be easily integrated into other systems, so as to provide the user with a more reliable and stable speech recognition result.
While the embodiment of the present invention have been illustrated and described, various modifications and improvements can be made by those skilled in the art. The embodiments of the present invention are therefore described in an illustrative but not restrictive sense. It is intended that the present invention may not be limited to the particular forms as illustrated, and that all modifications that maintain the spirit and scope of the present invention are within the scope as defined in the appended claims.

TABLE 2

the comparison between the word recognition rates of
MFCC and three compensation methods in the Aurora-2

Train-

Subway

Street

Subway

Babble

Car

Exhibition

Average

Restaurant

Street

Airport

static

Average

(M

(MIR

Average

(a) MFCC

Clean	98.93	99	98.96	99.2	99.0225	98.93	99	98.96	99.2	99.0225	99.14	98.97	99.055
20 dB	97.05	90.15	97.41	96.39	95.25	89.99	95.74	90.64	94.72	92.7725	93.46	95.13	94.295
15 dB	93.49	73.76	90.04	92.04	87.3325	76.24	88.45	77.01	83.65	81.3375	86.77	88.91	87.84
10 dB	78.72	49.43	67.01	75.66	67.705	54.77	67.11	53.86	60.29	59.0075	73.9	74.43	74.165
5 dB	52.16	26.81	34.09	44.83	39.4725	31.01	38.45	30.33	27.92	31.9275	51.27	49.21	50.24
0 dB	26.01	9.28	14.46	18.05	16.95	10.96	17.84	14.41	11.57	13.695	25.42	22.91	24.165
−5 dB	11.18	1.57	9.39	9.6	7.935	3.47	10.46	8.23	8.45	7.6525	11.82	11.15	11.485
Average	69.486	49.886	60.602	65.394	61.342	52.594	61.518	53.25	55.63	55.748	66.164	66.118	66.141

(b) CMS

Clean	98.93	99.09	99.02	99.04	99.02	98.93	99.09	99.02	99.04	99.02	99.08	99.06	99.07
20 dB	95.67	94.11	96.72	94.48	95.245	92.91	95.65	94.63	96.14	94.8325	95.52	96.1	95.81
15 dB	89.32	81.41	89.56	85.84	86.5325	80.56	88.39	85.36	87.2	85.3775	89.13	90.3	89.715
10 dB	68.96	57.07	67.94	64.05	64.505	61.22	66.17	66.33	66.21	64.9825	71.32	73.13	72.225
5 dB	38.56	28.48	34.95	31.04	33.2575	35.68	38.33	37.52	34.46	36.4975	38.47	44.95	41.71
0 dB	16.79	10.7	14.08	9.53	12.775	13.42	16.81	18.22	14.13	15.645	15.08	18.86	16.97
−5 dB	11.39	4.78	8.92	7.37	8.115	5.65	10.31	7.99	8.33	8.07	11.54	11.22	11.38
Average	61.86	54.354	60.65	56.988	58.463	56.758	61.07	60.412	59.628	59.467	61.904	64.668	63.286

(c) Prior art (U.S. Pat. No. 6,804,643 B1)

Clean	97.73	97.34	97.7	98.49	97.815	97.73	97.34	97.7	98.49	97.815	97.05	97.1	97.075
20 dB	92.69	92.41	93.53	90.96	92.3975	91.74	92.26	91.83	93.52	92.3375	86.34	89.51	87.925
15 dB	83.79	80.99	84.82	80.41	82.5025	80.78	83.62	81.15	82.32	81.9675	75.28	79.9	77.59
10 dB	66.99	60.4	62.87	62.02	63.07	60.39	63.39	60.39	60.04	61.0525	57.94	63.45	60.695
5 dB	42.77	31.47	32.03	35.98	35.5625	37.45	37.7	33.1	30.82	34.0175	35.62	41.17	38.395
0 dB	22.04	14.24	12.2	15.06	15.885	14.52	16.87	18.88	12.03	15.575	19.1	19.26	19.18
−5 dB	13.94	9.46	9.07	9.07	10.385	7.95	10.43	10.77	8.05	9.3	13.94	10.52	12.23
Average	61.656	55.902	57.09	56.886	57.8835	56.376	58.768	57.07	55.746	56.99	54.856	58.658	56.757

(d) The present invention

Clean	98.74	99	98.87	99.11	98.93	98.74	99	98.87	99.11	98.93	98.89	99.03	98.96
20 dB	96.87	95.22	97.2	95.19	96.12	94.47	96.7	96.15	96.7	96.005	96.1	96.67	96.385
15 dB	93.21	84.98	93.11	90.19	90.3725	84.89	90.99	89.83	89.51	88.805	92.26	93.17	92.715
10 dB	77.74	62.03	73.64	71.8	71.3025	64.54	72.34	70.18	71.18	69.56	79.46	80.47	79.965
5 dB	46.91	31.62	37.16	38.66	38.5875	37.89	41.66	39.9	37.15	39.15	52.29	51.03	51.66
0 dB	20.97	13.03	12.29	13.48	14.9425	16.12	17.2	18.76	11.94	16.005	21.52	21.64	21.58
−5 dB	11.27	6.32	8.92	8.42	8.7325	7.03	10.61	9.13	7.25	8.505	12.25	10.52	11.385
Average	67.14	57.376	62.68	61.864	62.265	59.582	63.778	62.964	61.296	61.905	68.326	68.596	68.461

indicates data missing or illegible when filed

Claims

1. A speech recognition system with cepstral noise subtraction, comprising:

a filterbank energy extractor, for obtaining a plurality of first feature vectors according to a voice signal;

a cepstral noise subtraction device, for obtaining a first feature vector of a preset voice frame and first feature vectors of a plurality of voice frames before the preset voice frame, so as to calculate a feature mean vector, and calculate a second feature vector of a preset voice frame according to the first feature vector, the feature mean vector, a first scalar coefficient, and a second scalar coefficient of the preset voice frame;

a cepstral converter, for converting the second feature vector of the preset voice frame into a cepstral feature vector;

a model trainer, for calculating a model parameter according to the cepstral feature vector; and

a speech recognizer, for calculating a recognized voice signal according to the cepstral feature vector and the model parameter.

2. The speech recognition system according to claim 1, wherein the cepstral noise subtraction device comprises:

a feature mean vector calculator device, for obtaining the first feature vector of the preset voice frame and the first feature vectors of the plurality of voice frames before the preset voice frame, so as to calculate the feature mean vector;

a first multiplier, for multiplying the feature mean vector by a negative value of the first scalar coefficient, so as to calculate a first multiplication result;

a first adder, for adding the first feature vector of the preset voice frame with the first multiplication result, so as to calculate an addition result;

a second multiplier, for multiplying the first feature vector of the preset voice frame by the second scalar coefficient, so as to calculate a second multiplication result;

a comparator, for comparing whether the addition result is greater than the second multiplication result, and outputting a control signal; and

a multiplexer, for switching the second feature vector of the preset voice frame into the addition result or the second multiplication result according to the control signal.

3. The speech recognition system according to claim 2, wherein when the addition result is greater than the second multiplication result, the second feature vector of the preset voice frame is the addition result, and when the addition result is smaller than the second multiplication result, the second feature vector of the preset voice frame is the second multiplication result.

4. The speech recognition system according to claim 3, wherein the first scalar coefficient is between 0.01 and 0.99, and the second scalar coefficient is between 0.01 and 0.99.

5. The speech recognition system according to claim 2, wherein the feature mean vector calculator device comprises:

a plurality of delayers, each delaying a unit of time to obtain the first feature vectors of the plurality of voice frames before the preset voice frame;

a second adder, for summing the first feature vectors, so as to calculate a sum of the first feature vectors; and

a third multiplier, for multiplying the sum of the first feature vectors by a reciprocal of the number of the voice frames, so as to calculate the feature mean vector.

6. The speech recognition system according to claim 2, wherein the feature mean vector calculator device calculates the feature mean vector through mean calculation methods including geometric mean, median, mode, or norm.

7. The speech recognition system according to claim 1, wherein a number of the plurality of voice frames before the preset voice frame is between 2 and the total number of voice frames of a sentence.

8. The speech recognition system according to claim 1, further comprising a differential operator for calculating a first-order difference, or a first-order difference and a second-order difference, or a first-order difference to a high-order difference of the cepstral feature vector.

9. The speech recognition system according to claim 1, wherein the filterbank energy extractor is a log Mel filterbank energy extractor.

10. The speech recognition system according to claim 9, wherein the cepstral converter is a discrete cosine transformer.

11. A speech recognition method with cepstral noise subtraction, comprising:

obtaining a plurality of first feature vectors according to a voice signal;

obtaining a first feature vector of a preset voice frame and first feature vectors of a plurality of voice frames before the preset voice frame, so as to calculate a feature mean vector;

calculating a second feature vector of a preset voice frame according to the first feature vector, the feature mean vector, a first scalar coefficient, and a second scalar coefficient of the preset voice frame;

converting the second feature vector of the preset voice frame into a cepstral feature vector;

calculating a model parameter according to the cepstral feature vector; and

calculating a recognized voice signal according to the cepstral feature vector and the model parameter.

12. The speech recognition method according to claim 11, wherein the step of calculating a second feature vector of the preset voice frame further comprises:

obtaining the first feature vector of the preset voice frame and first feature vectors of the plurality of voice frames before the preset voice frame, so as to calculate the feature mean vector;

multiplying the feature mean vector by a negative value of the first scalar coefficient, so as to calculate a first multiplication result;

adding the first feature vector of the preset voice frame with the first multiplication result, so as to calculate an addition result;

multiplying the first feature vector of the preset voice frame with the second scalar coefficient, so as to calculate a second multiplication result;

comparing whether the addition result is greater than the second multiplication result, and outputting a control signal; and

switching the second feature vector of the preset voice frame into the addition result or the second multiplication result according to the control signal.

13. The speech recognition method according to claim 12, wherein when the addition result is greater than the second multiplication result, the second feature vector of the preset voice frame is the addition result, and when the addition result is smaller than the second multiplication result, the second feature vector of the preset voice frame is the second multiplication result.

14. The speech recognition method according to claim 11, wherein the step of calculating a feature mean vector further comprises:

using a plurality of delayers each delaying a unit of time to obtain the first feature vectors of the plurality of voice frames before the preset voice frame;

summing the first feature vectors to calculate a sum of the first feature vectors; and

multiplying the sum of the first feature vectors with a reciprocal of the number of the voice frames, so as to calculate the feature mean vector.

15. The speech recognition method according to claim 11, wherein the feature mean vector is calculated through mean calculation methods including geometric mean, median, mode, or norm.

16. The speech recognition method according to claim 11, further comprising a difference operation step, for calculating a first-order difference, or a first-order difference and a second-order difference, or a first-order difference to a high-order difference of the cepstral feature vector.

17. The speech recognition method according to claim 11, wherein the first feature vectors are log Mel filterbank energy feature vectors.

18. The speech recognition method according to claim 11, wherein the cepstral feature vector is a Mel cepstral feature vector.